BetselotYitagesu/Fine-tuning-LLM-for-Amharic-NER

✅ Task 1 & 2: Data Collection, Cleaning & Manual NER Annotation 📦 Task 1: Data Ingestion & Preprocessing We collected Amharic e-commerce messages from 5 public Telegram channels using the Telethon Python library. These messages often include product names, prices, and store locations. 📡 Channels Scraped Channel Name Messages ShagerOnlineStore 89 Leyueqa 112 marakibrand 46 qnashcom 160 MerttEka 102 🔧 Preprocessing: Removed URLs, mentions, emojis, and non-Amharic characters. Preserved important text like numbers, prices, and Amharic punctuation. Output a cleaned message column for further annotation. 📂 Output Files Raw data: data/raw/*.csv Cleaned data: data/processed/*_clean.csv Scripts used: Scraper: scr/ingestion/scrape_telegram.py Cleaner: scr/preprocessing/preprocess_text.py 📝 Task 2: Manual NER Labeling in CoNLL Format We manually labeled 30–50 cleaned messages using the standard CoNLL format, focusing on 3 key entity types for NER fine-tuning: 🏷️ Entity Types and Labels Entity Tags Product B-Product, I-Product Price B-PRICE, I-PRICE Location B-LOC, I-LOC Other O 🧰 Tool Used: Custom Python annotation interface: scr/annotation/amharic_ner_annotator.py Prompts the user to label each token manually. Saves annotations in: amharic_ner_annotations.conll 📄 CoNLL Format Example: Skechers B-Product archfit I-Product Price B-PRICE 3400 I-PRICE birr I-PRICE አድራሻ B-LOC መክሲኮ I-LOC ask 3: Model Evaluation Evaluated the fine-tuned XLM-RoBERTa NER model using the seqeval metric for entity-level precision, recall, and F1-score. Implemented evaluation code to map predicted token labels back to entities and compared with ground truth. Generated detailed classification reports to identify model strengths and weaknesses on PRODUCT, PRICE, and LOCATION entities. Evaluated on test splits of the dataset and saved performance metrics for reference. Task 4: Model Inference Developed an inference pipeline to predict named entities on new Amharic Telegram messages. …